In which languages can an AI concierge reply?

In the languages you release. Technically, common European languages are no problem, because the stored knowledge is output in translation. It makes sense to limit this to the languages in which your front desk can continue the handover. Otherwise the system may reply in Hungarian and the follow-up come in German.

What happens if the concierge gives wrong information?

Responsibility stays with the property, just as with a temp at the front desk. That is why price, cancellation and health questions belong in the handover to a person from the outset. In addition, all conversations should be logged, so you can trace a wrong answer and correct the knowledge base instead of guessing.

Does an AI concierge replace the night reception?

No. At night it answers questions that have a fixed answer and records requests. It does not open a door, arrange a room change or deal with a water leak. Properties without night staff gain calm in their inbox, but still need a reachable phone number for emergencies.

How long does set-up in the property take?

The larger part is editorial rather than technical: gathering the property's knowledge and agreeing on it. Plan half to a full working day for this, split over two appointments, plus a test phase of one to two weeks in which every answer is proofread before the concierge replies without supervision.

AI concierge for hotels: what it handles, what it does not

Key takeaways

Four topic blocks make up most pre-arrival messages in most properties: arrival and parking, times, pets, and in-room facilities.
Price reductions, complaints, medical questions and everything contractual go straight to a person.
The handover needs three things: a threshold, a channel and a record that the person sees before replying.
Without maintained knowledge, a concierge replies in a friendly way and gets things wrong; keeping the knowledge current is work for the property, not for the software.
Article 50 of the EU AI Act (KI-Verordnung) requires that guests can recognise they are writing with an AI system.

An AI concierge answers the questions that are asked again every day in every property: how do I get there, where do I park, when is breakfast, may the dog come along. For matters of money, goodwill, health or contracts, it does not answer but passes them on to a person.

This boundary is the real design. Draw it cleanly and you get a system that quietens the inbox. Leave it open and you get a system that at some point confirms a price nobody agreed to.

Questions a concierge can answer safely

In most properties the messages before arrival fall into a few blocks. The first is arrival itself: address, last turn, railway station, bus, winter tyre requirement on the access road. The second is parking: how many spaces, whether they cost anything, whether there is a charging point, whether a van fits in the garage. The third is times: check-in window, late arrival, breakfast on weekdays and at weekends, kitchen closing time, quiet hours. The fourth is pets and facilities: is a dog allowed, in which rooms, at what fee, is there a cot, a kitchenette, air conditioning, a hairdryer.

All of these questions have an answer that is fixed in the property and rarely changes. That is exactly what a concierge is suited to: it gives the answer in full sentences, in the guest's language, at 23:00 just as at 7:00, and it does not grow impatient the twentieth time.

Questions that belong with a person

Some messages are harmed by an automated answer even when its content is correct. Price reductions and special terms are among them: whoever asks for a better price is negotiating, and negotiating is for the property to do. Complaints belong here too, even small ones. A complaint is an offer to talk, and a machine does not take it up. Medical and dietary questions belong here as well, from nut allergies to a step-free bathroom, because an imprecise answer is a real risk for the guest.

Add everything contractual: cancellation terms in individual cases, deposits, group contracts, invoices to companies, registration forms and identity obligations. And finally the requests that are not questions at all, such as the guest who writes that they will arrive late because something happened on the train. A person answers that.

Working through the figures for an aparthotel with 18 units

A worked example with stated assumptions, not a survey: an aparthotel with 18 units, 65 percent occupancy across the year, three nights per stay, two messages before each arrival, and two minutes per message for reading, checking the occupancy plan and replying. The number of messages is the least certain value; count for one week and you will have your own figure.

Occupied nights: 18 × 365 × 0.65, around 4,270
Stays at three nights each: around 1,420
Messages before arrival, two per stay: around 2,840
Time at two minutes per message: around 95 hours a year
If the concierge fully answers half of them: around 47 hours a year stay free for the front desk

What is left are the messages that deserve attention anyway. The gain is not in staff reduction. It lies in the front desk no longer having to search for the difficult messages among twenty simple ones.

How the handover to the front desk works

A handover needs three decisions. First, a threshold: the concierge replies only when it finds the answer in the stored knowledge. If it does not find it, or only partly, it says so and passes the message on. An honest "I will look that up for you, and the front desk will get back to you by 9:00" is better than a guessed answer.

Second, a channel: the forwarded message lands where the property already works, in the shared inbox, in the booking system or in the channel for guest communication via WhatsApp. Not in a second tool that nobody opens. Which questions suit WhatsApp and which are better left with the front desk is shown in the overview of WhatsApp in guest communication.

Third, a record: whoever takes over the message sees the earlier conversation, the booking number and the arrival date. Without this context the conversation starts again from the beginning, and the guest has to explain their request a second time. A well-kept dataset on guests and stays, as provided by a guest data system, is the prerequisite here.

What you have to supply yourself: the property's knowledge

A concierge has no knowledge of its own about your property. It only has what you store. In practice that means room categories with sizes and beds, breakfast times split by season, parking spaces with number and price, the pet rule with its fee, directions from two directions, facilities per category, cancellation and deposit rules worded exactly, and contact persons with their hours.

This collection must belong to someone. In properties where this works, one person owns it and touches it with every change: a new breakfast time in winter, roadworks on the access road, a lift out of service. Allow half a day for the first capture and a few minutes per change afterwards. If the collection is not kept up, the concierge still answers politely, only with last summer's times.

The guest must be able to tell they are writing to a machine

The EU AI Act, Regulation (EU) 2024/1689, sets out transparency duties in Article 50 for systems that interact directly with people. Those affected should be able to recognise that they are dealing with an AI system, unless this is already obvious [Q1] [Q2]. For a property, this simply means that the concierge identifies itself rather than presenting itself as a member of staff.

How this must look in the individual case, and which further duties apply to you, is a legal question; have it checked by a specialist. For Austria, RTR runs an AI service desk (KI-Servicestelle) as a point of contact for questions about the regulation [Q3]. An in-depth classification for accommodation businesses is available under the EU AI Act for hospitality.

How to tell after three months whether it works

Four key figures are enough, and all four are in the system without anyone having to count. First, the share of conversations that ended without a handover; this is the real relief. Second, the number of handovers that arose outside front desk hours; it shows whether the concierge creates night work or prevents it. Third, the number of corrections to the knowledge base per month: many are normal in the first weeks, and after the second month there should be few. Fourth, complaints about an answer, each of which should be read rather than counted.

What you will not find in these figures is an effect on booking volume. A concierge answers questions; whether they turn into a booking depends on price, room and date. Anyone who justifies the set-up with an expected revenue effect will be disappointed. Anyone who justifies it with quieter evenings and faster answers will see the result in the first month.

When an AI concierge is not worth it

For very small properties the arithmetic rarely works. Anyone who rents out four holiday apartments and receives fifteen messages a week saves only a few hours a year, and pays for set-up and upkeep. Here a well-built information section on your own site, plus a set of template replies on the phone, is usually the better answer.

Nor is it worth it if the knowledge exists only in people's heads and nobody has time to write it down. And not if guests explicitly want personal contact. At chalets with all-round service, a quick answer at midnight is not an advantage but a break in the promise.

If you want to know which of your messages could be automated at all, count for one week and sort them by the four blocks above. On that basis, an AI concierge for your property can be tailored sensibly. Book a call.

Sources and further reading

[Q1] Regulation (EU) 2024/1689 (EU AI Act), EUR-Lex

[Q2] Article 50: transparency obligations (text version)

[Q3] RTR: AI service desk (KI-Servicestelle)

What an AI concierge takes off your hands in hotels